Semantic Segmentation of Remote Sensing Image Based on Encoder-Decoder Convolutional Neural Network

Zhehan Zhang; Wei Fang; Lili Du; Yanli Qiao; Dongying Zhang; Guoshen Ding

doi:10.3788/AOS202040.0310001

Acta Optica Sinica, Volume. 40, Issue 3, 0310001(2020)

Semantic Segmentation of Remote Sensing Image Based on Encoder-Decoder Convolutional Neural Network

Zhehan Zhang^1,2, Wei Fang^1、*, Lili Du¹, Yanli Qiao¹, Dongying Zhang¹, and Guoshen Ding^1,2

¹Key Laboratory of Optical Calibration and Characterization, Anhui Institute of Optics and Fine Mechanics, Chinese Academy of Sciences, Hefei, Anhui 230031, China

²University of Science and Technology of China, Hefei, Anhui 230026, China

show less

Abstract Get PDF(in Chinese)

Figures & Tables(11)

Fig. 1. Architecture comparison. (a) U-Net; (b) SegNet; (c) SegProNet

Download full size

Fig. 2. Remote sensing image semantic segmentation network structure of SegProNet

Download full size

Fig. 3. Maxpooling indices and upsampling

Download full size

Fig. 4. Display of data sets. (a) Training set image; (b) corresponding label visualization image

Download full size

Fig. 5. Some training set images

Download full size

Fig. 6. Comparison of experimental results. (a) Original image; (b) original label visualization result; (c) U-Net segmentation result; (d) SegNet segmentation result; (e) segmentation result of SegProNet+ReLU; (f) segmentation result of SegProNet+ELU

Download full size

Fig. 7. Comparison of experimental details. (a) Original label visualization result; (b) U-Net segmentation result; (c) SegNet segmentation result; (d) segmentation result of SegProNet+ReLU; (e) segmentation result of SegProNet+ELU

Download full size

Fig. 8. Each network loss and accuracy curves. (a) U-Net; (b) SegNet; (c) SegProNet; (d) SegProNet+ELU

Download full size

Table 1. Information of labels
View table
Table 1. Information of labels
Feature category No. Label color (R,G,B)
Background 0 (0,0,0)
Vegetation 1 (50,205,50)
Building 2 (245, 254, 0)
Water 3 (0, 255, 255)
Road 4 (255, 92, 75)

Table 2. Comparison of training and prediction time of each network
View table
Table 2. Comparison of training and prediction time of each network
Network category Training time /h Prediction time /h
U-Net 16.3 0.26
SegNet 27.6 0.61
SegProNet 23.2 0.52
SegProNet+ELU 22.6 0.51

Table 3. Evaluation indicators of each method

View table

Table 3. Evaluation indicators of each method

Category	Evaluation	Vegetation	Building	Water	Road
U-Net	Precision	0.7565	0.5232	0.7368	0.6904
	Recall	0.6341	0.7847	0.8237	0.7571
	IoU	0.5266	0.4575	0.6364	0.5652
SegNet	Precision	0.8340	0.7254	0.8555	0.7576
	Recall	0.8164	0.8338	0.8541	0.8722
	IoU	0.7023	0.6339	0.7464	0.6819
SegProNet	Precision	0.8533	0.7508	0.8837	0.7815
	Recall	0.8239	0.8587	0.8653	0.8736
	IoU	0.7216	0.6682	0.7768	0.7021
SegProNet+ELU	Precision	0.8531	0.8861	0.8792	0.8664
	Recall	0.8534	0.7528	0.8914	0.8164
	IoU	0.7441	0.6863	0.7942	0.7251

Tools

Get Citation

Copy Citation Text

Zhehan Zhang, Wei Fang, Lili Du, Yanli Qiao, Dongying Zhang, Guoshen Ding. Semantic Segmentation of Remote Sensing Image Based on Encoder-Decoder Convolutional Neural Network[J]. Acta Optica Sinica, 2020, 40(3): 0310001

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites